7th int. workshop applications in electronics pervading ... · unibo contribution to epi ......
TRANSCRIPT
7th Int. Workshop
Applications in Electronics Pervading Industry, Environment & Society
APPLEPIES 201911-13 September, School of Engineering, University of Pisa, www.applepies.eu
Round Table:
Trend in Italy on EuroHPC & the European Processor Initiative
CINECA E4 STMicroelectronics UniBologna UniPisa
2021 – H2
PCIe board (WS compatible)
with < RHEA β version
ARM MPPA
eFPGA EPAC
HBM
memories
DDR
memories
PCIe gen5
linksHSL
links
D2D links
to adjacent
chiplets
EPAC
STX
VPU
VRP
Bridge to GPP
Bridge to GPP
UNIBO DEI EEES laboratory research excellence is on HW and SW co-design of energy-efficient computing system.
The HPC team’s research focuses on methods and tools for monitoring and power management of large-scale computing systems
The Compilers & Runtime team’s research focuses on the programming support for heterogenous computing systems
The Digital Design team’s research focuses on the design of cutting edge computing architectures and their silicon implementation
based on the parallel-ultra-low power (PULP) open-source hardware platform
The Edge AI team’s research focuses on the development of AI-enabled IoT and smart sensors systems and applications
UNIBO contribution to EPI
• UNIBO activities focus on the design of the HW and SW power management subsystem of the EPI processors.
• UNIBO activities focus on the design of the HW and programming support for offloading and supporting computation on the EPI
specialized accelerators.
CONTRIBUTION TO EPI
Power Controller
VRM
BMCPE
S
Operating System
Application
System Management / RM
GovernorsIn band
Hints/Prescription
Power Cap Energy vs. Througput
DIMM
RJ45
Syste
m M
anagem
ent
/ RM
Out of bandNode Power Cap
RAS
Main Archi. Blocks w. :- Sensors (PVT, Util, archi)- Controls (f,Vdd,Vbb,PG,CG)
- In band a.k.a low latency / user-space telemetry (power, perf, …)- O.S. PM governors:
- cpufreq/ cpuidle- Based on O.S. metrics- Slow & often unused
- Low latency PM requests and/or suggestions
- From the Application/run-time
- Power cap => Max perf @ P<Pmax- Energy => Min Energy @ f=f*- Throughput => F > Fmax @ T,P<Max
- Out-of-band – zero overhead telemetry
- Node Pcap – Max perf @ Pnode<Pmax- RAS – error and conditions reporting
EPI
chip
24
Private consortium under public control:
• 67 universities
• 9 public research entities
• 3 university hospitals
• MIUR Our numbers:
100M€ turnover
21th in the TOP500
800 employees
in BO, MI, Roma, NA
Main activities• Supercomputing for research• IT Support to MIUR• IT Support to universities• Technology transfer
•
•
•
•
•
•
25
CONTRIBUTION TO EPI
26
Japan (2022)
U.S. (2021-2023)
China (2021-2022)Xiaomi, ShenWei, Hygon
EuroHPC (2023)
Aurora 21 (Cray/Intel + GPUs)
Frontier (Cray/Intel + GPUs)
Fugaku (Fujitsu/ARM)
Hygon x86 (AMD-Chinese Joint Venture) + Manycore
accelerators (Matrix-2000+?)
27
2009 2012 2016 2019/2020 2021 2023-2025 2025-2027
IBM SP6
Power6
Fermi
IBM BGQ
PowerA2
Marconi
Lenovo
Xeon+KNL
Marconi
PPI4HPC
ICEI - PPIHBP
Leonardo
Pre-exascale with
EuroHPC
contribution
Post-Marconi
Exascale pilot
technology
Exascale with EuroHPC
Procurement in
progress
100TF1MW
2PF1MW
11PF+9PF
3.5MW
50PF+20PF
~4MW
20x
5x
5x
1x
(latency cores)
Paradigm
change
10x
(in total)
> 250 PF> 20PF
~10MW
EuroHPC
Tier-1Exascale
PilotContainer orchestration
(IP over low latency
network)
1ExaF~20MW
EuroHPCConverging architecture
(containerized + traditional
HPC schedulers)
28
PUE < 1.1ECMWF
Computer Rooms:
10 MW (2020) → 20 MW
(2023)
250-300 petaFLOPS
Estimated to end up in the
top 3 spots of Top500
E4 Computer Engineering
Since 2002, E4 Computer Engineering has been innovating and actively encouragingthe adoption of new computing and storage technologies. Because new ideas are soimportant, we invest heavily in research and hence in our future. Thanks to ourcomprehensive range of hardware, software and services, we are able to offer ourcustomers complete solutions for their most demanding workloads in: HPC, Big-Data,AI, Deep Learning, Data Analytics, Cognitive Computing and for any challengingStorage and Computing requirements.
29
Silver Levelhttps://openpowerfoundation.org/Cosimo Gianfreda
Member of the Steering Board http://www.etp4hpc.eu
Member of the Consortiumhttp://european-processor-initiative.com
Member of CERN openlabhttps://openlab.cern/
https://www.openfoam.com/
Member of HiPEAC
https://www.hipeac.net/
Member of the OEHI (Open Edge and HPC Initiative)http://www.open-edge-hpc-initiative.org/
Member of the MaXCenter of Excellencehttp://www.max-centre.eu/
E4 Computer Engineering
E4 Computer Engineering
2021 – H2
PCIe board with < RHEA β version
CONTRIBUTION TO EPI
• Technological leadership in a key sector
• Create a European ecosystem for microelectronics:
• Engineers, IP’s, foundries, technologies
• Create a European ecosystem for exascale-class applications
• Create an Italian ecosystem of expertise & knowledge
• We have to retain our talents
• Drive the completeness of the Italian and European industry
• Our industrial system faces a worldwide competition
The value of EPI for the industry
STMicroelectronics is a key player in the Server & Data Center Power Management arena and brings innovation by:
• Developing new ICs/IPs to support the advanced power management features of the newest CPU, DDR and ASIC chipset
ensuring best in class fast load transient response, high precision voltage positioning and accurate telemetry;
• Enabling new Architecture & Applications to achieve higher conversion efficiency and density required by the modern power
Server and Data Center services.
STMicroelectronics contribution to EPI
• ST activities focus on the development of a new generation of power conversion ICs dedicated to 48V power architecture
in order to significantly reduce the total cost of ownership in ultra-high-efficiency HPC based on EPI processor;
• ST enables also the development of the advanced power management for the EPI processor which is based on the adoption of
STMicroelectronics Power Management ICs.
ST CONTRIBUTION TO EPI
ST 48V SERVER & DATA CENTER
POWER MANAGEMENT
STC (unregulated)48V
MU
LTIP
HA
SE D
IGIT
AL
CO
NTR
OLL
ER
STBuck(regulated)
Isolated ResonantArchitecture
2° Stage (12V input)
Single Stage
SPS
SPS
SPS
SPS
SPS
SPS
SPS
SPSDu
al Sta
ge
Co
nv
ers
ion
Dire
ct
Co
nv
ers
ion
1° Stage (IBC*)*Intermediate Bus Conversion
SPS = Smart Power Stage
12V
CPU/DDR/ASIC
48V
Powering Server & Modern Data Center via the highest efficiency & density solutions
Complete solutions for Direct or Dual Stage CPU / DDR / ASIC power management
in 48V ecosystem
CPU/DDR/ASIC
Safety Critical
Industry 4.0
& Robotics
AeroSpace
Automotive
Servers
& Cloud
HPCCore
Drivers
sovereignty
AI
accelerators
AI accelerators (Posits)
DNN & ML and crypto benchmarking on ARM 64b
On-chip TRNG & CSPRNG
Crypto accelerators for hashing, symmetric key and public key crypto
Vertical chain for Secure OTA update of FW/SW
Safety & security of the MCU – HPC monitoring link
CONTRIBUTION TO EPI
1st EPI chip
Crypto Accelerators & on-chip TRNG/CSPRNG
TRNG (FiRo/GaRo oscillator based) & CSPRNG
Crypto accelerators, up to 300 Gbps real-time encryption/decryption
in 7 nm (needed to exploit HBM)
Configurable AES 126b/256b keys with 9 chipper modes
Configurable SHA2/SHA3 up to 512b hash functions
ECC-based public key crypto accelerators (ECDH, ECSIE,
ECDSA,..)
Open SSL Benchmarks on multi-core 64b ARM
X 3 order of magnitude gain thanks to crypto accelerators
Crypto accelerator ad-hoc policies integrated on-chip
Keys storage and management
Crypto configuration ad-hoc policy
Secure channels
V2X
Embedded
HPCV2X
V2X channel security & 5G on EPI
Over the Air SW/FW Update - Vertical chain
Post-quantum resistant
App SW (attributed based)
eLinux OS coexisting with RTOS
Hypervisor
EPI HW with crypto accelerators
5G resource control & adaptive allocation
spectrum, beam-forming/steering, channels,
power, goodput (modulation, coding,..)
Data Type (tot_bits, exp_bits) Accuracy on 10,000 images
Float32 98,88%
Posit16,2 98,88%
Posit14,2 98,85%
Posit12,2 98,66%
Posit10,0 98,69%
Posit8,0 97,24%
Posit Processing Unit
vs FPUMachine Learning test
(K-NN & A-NN classifiers)
Same or better accuracy with
halved bit-size
Posit Processing Unit vs FPUDNN tests
(MNIST, CIFAR, CoCo, Yolo, Alexnet,
Imagenet, GTRSB,…)
Posit has similar accuracy of Float with
complexity reduction x4
Float16
Posit8
(exp: 1bit)
Float32
Posit16
(exp: 3bit)
Posit Processing Unit available as
SW-library or IP macrocell
(HLS on FPGA or 7nm IC)
Float to Posit conversion
for data storage saving
Posit as compressed float
Accelerators for AI computing
OPEN DISCUSSION
Trend in Italy on EuroHPC & the European
Processor Initiative
Round Table:
Trend in Italy on EuroHPC & the European Processor Initiative
Federico Ficarelli, CINECA
Fabrizio Magugliani, E4
Alberto Rossi, STMicroelectronics
Andrea Bartolini, UniBologna
Sergio Saponara, UniPisa